DeepValue: A Comparable Framework for Value-Based Strategy by Machine Learning

Value relevant analysis is one of the key stock trading strategies of stock investment which is based on financial statement that represents the intrinsic investing value of firms. Human analysts have dominated the value interpretation of companies so far, in spite of the numerous efforts being made by machine learning researchers nowadays. The complexity of hundreds of accounting terms in financial statements and the latent interaction among industrial contextual factors are hard to be integrated into machine learning approaches. The analysis of profitability and potential for specific companies are often unique and not applicable to the others. In this paper, we constitute a unified learning framework, named as DeepValue, for knowledge transfer from and value extraction of collective financial status for value-based investment strategy. We validate four kinds of feature set from 11 years of financial statements of 90 semiconductor companies in TWSE. It incorporates deep Learning surrogate function for feature mapping, Multitask Learning (MTL) for knowledge transfer and feature sharing, and Long Short-Term Memory (LSTM) for value extraction. This value is then used to assess the degree of value-price gaps for stock selection. Experiment results demonstrate that the proposed framework is able to quantify the potential of stocks. The derived recommendation lists significantly outperform market and EPS-based selection.